On October 15, Microsoft gave enterprise agents an identity of their own. In a Microsoft Copilot Blog announcement, the company introduced Copilot Studio agent accounts, business process memory and triggers. Agent identity is no longer a borrowed login. It is a credential, a mailbox and a memory that persist after the person who created the agent has logged off.
Microsoft says agents with accounts work in Microsoft 365 "with their own credentials, similar to how any person would." People can invoke the agent in Teams and Outlook, assign it work by email, and @mention it in Word, Excel and PowerPoint comments. The agent can work with other agents and tools "using their own approved credentials and scoped permissions, without relying on any individual user's or maker's credentials," and its actions are attributed to the agent itself.
What agent identity changes in practice
Three things change at once. First, attribution. Microsoft says accounts make it clearer "which actions were taken by agents and which were taken by people," while administrators can scope access and manage the agent lifecycle. Second, persistence. Triggers in the GitHub Copilot harness let an agent start a cycle when an email arrives or a set cadence falls due, with "no additional prompt required." Third, reach. Microsoft names Ooredoo Qatar as an early customer, where an agent works across Outlook, Teams and Word to review business cases, apply company policy and precedents, and prepare briefs for executive review.
The preview has limits. As NeoTeo's summary of Microsoft's documentation notes, an account-based agent gets its own Microsoft Entra ID and mailbox after publication, but it cannot respond to people outside its tenant or read documents protected by labels such as "Confidential" or "Internal Only" in the preview, and email responses are limited to conversations inside the tenant with up to 250 people. An account is an identity, not a blank check. Makers still configure the permissions for each tool.
The identity data was already warning about this
The case for giving agents a governed identity is strong because the current state is weak. SpyCloud's 2026 Identity Threat Report, based on a survey of 750 cybersecurity leaders and practitioners at organisations with 500 or more employees, found compromised non-human identities were the primary entry point in 31 percent of cases, nearly twice phishing and social engineering at 17 percent. Non-human identity misuse was the most common identity event at 42 percent. Yet 95 percent of organisations believed they had adequate visibility, while only 36 percent were actually monitoring those identities.
BlueRadius's AI Agent Security Report 2026, which coded 21 publicly documented agent security cases, found excessive privilege in 76 percent, prompt injection in 71 percent, and a bypass or absence of human approval in every case. A separate survey reported by Cybersecurity Insiders from Akeyless, based on 400 IT and security leaders, found two-thirds of organisations believed their agents had already accessed data outside their intended scope. The same reporting cites a Cloud Security Alliance study finding two-thirds could not clearly distinguish agent actions from human actions in audit logs.
Read together, those numbers describe the problem agent accounts are meant to fix. An agent that borrows a person's credentials inherits that person's reach and blurs the log. An agent with its own scoped identity can, in principle, hold less access and leave a cleaner record. The principle holds only if the scope is narrow, the sponsor is named, and a person can revoke the identity.
A shared memory with no privacy boundary
Memory is where the preview asks the hardest question. Microsoft describes business process memory as carrying forward "users' preferences and business-process context into each new cycle," so that a point one person raises "becomes a shared resource for everyone else to benefit from." That is the value. It is also the exposure.
As the NeoTeo summary puts it, memory for an agent with its own account "is shared among the people who work with it," and Microsoft says the preview has "no user privacy boundary for that memory and no option to opt out." Microsoft's separate general memory works differently, with a private store for each user that makers can toggle. The account-based design chooses the team record over the private one. A company adopting it needs to decide, before the first cycle, what should never enter that shared memory.
That decision is the same one raised in our earlier reporting on the Microsoft team knowledge base post from October 18, 2026 and in Memory Poisoning: The Attack That Turns AI Agent Memory Into a False Record. A shared record helps only while someone keeps it true, scoped and attributable.
Jilu takes the narrower position on the same problem, set out in How Jilu Works. An agent should hold no access apart from the person who invoked it. Each action is logged against that person. Knowledge the agent drafts stays a draft until a person reviews and publishes it. That model gives up some of the autonomy an agent account offers. It keeps one thing the new accounts still have to prove in production: when the agent acts, a named person remains answerable for what it read, what it remembered and what it did.
The direction Microsoft set on October 15 is likely to hold. Agents will have identities. The open question is not whether an agent gets an account. It is whether the account comes with a named owner, a narrow scope, a memory a person has reviewed, and a log that separates the agent's decision from the person's intent. Companies that settle those four points first will get the persistence without inheriting the standing invitation that SpyCloud describes: an identity no one owns, usable until someone notices.